Semi-Supervised Intracranial Aneurysm Segmentation from CTA Images via Weight-Perceptual Self-Ensembling Model

Segmentation of intracranial aneurysm (IA) from computed tomography angiography (CTA) images is of significant importance for quantitative assessment of IA and further surgical treatment. Manual segmentation of IA is a labor-intensive, time-consuming job and suffers from inter- and intra-observer variabilities. Training deep neural networks usually requires a large amount of labeled data, while annotating data is very time-consuming for the IA segmentation task. This paper presents a novel weight-perceptual self-ensembling model for semi-supervised IA segmentation, which employs unlabeled data by encouraging the predictions of given perturbed input samples to be consistent. Considering that the quality of consistency targets is not comparable to each other, we introduce a novel sample weight perception module to quantify the quality of different consistency targets. Our proposed module can be used to evaluate the contributions of unlabeled samples during training to force the network to focus on those well-predicted samples. We have conducted both horizontal and vertical comparisons on the clinical intracranial aneurysm CTA image dataset. Experimental results show that our proposed method can improve at least 3% Dice coefficient over the fully-supervised baseline, and at least 1.7% over other state-of-the-art semi-supervised methods.